Discrimination of Handwritten and Printed Texts Based on Frame Features and Viterbi Decoder

Qin Lin; Junfeng Xia; Zhengzheng Tu; Yutang Guo

doi:10.3788/LOP56.061003

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 6 >Page 061003 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 6, 061003 (2019)

Discrimination of Handwritten and Printed Texts Based on Frame Features and Viterbi Decoder

Qin Lin^1、*, Junfeng Xia², Zhengzheng Tu², and Yutang Guo¹

Author Affiliations

¹ School of Computer Science Technology, Hefei Normal University, Hefei, Anhui 230601, China

² College of Computer Science and Technology, Anhui University, Hefei, Anhui 230039, China

show less

DOI: 10.3788/LOP56.061003 Cite this Article Set citation alerts

Qin Lin, Junfeng Xia, Zhengzheng Tu, Yutang Guo. Discrimination of Handwritten and Printed Texts Based on Frame Features and Viterbi Decoder[J]. Laser & Optoelectronics Progress, 2019, 56(6): 061003 Copy Citation Text

show less

Fig. 1. Flow chart of proposed algorithm

Download full size

Fig. 2. Schematic of hidden Markov model

Download full size

Fig. 3. Schematic of state transition of hidden Markov model

Download full size

Fig. 4. All possible Viterbi decoding paths

Download full size

Fig. 5. Discrimination results of handwritten and printed texts. (a) Frame feature decoding results mapped to text line images; (b) longitudinal image segmentation; (c) re-determination results in each region

Download full size

Layer name	Output size	Convolution kernel
conv1-1conv1-2conv1-3pool1	24×[(W-3)/2+1]	32@3×3, d_pad=164@1×164@3×3, d_pad=13×3 Max pool, e_stride=2
conv2-1conv2-2conv2-3conv2-4pool2	6×[(W-3)/2-3]	64@1×1128@3×3, e_{stride_h}=264@1×1128@3×3, d_pad=13×3 Max pool, e_{stride_h}=2
conv3-1conv3-2fc	1×[(W-3)/2-3]	256@3×1, d_pad=1128@3×1S@1×1

Table 1. Convolutional neural network structure of OCR based on text line

Method	HandwrittenAccuracy	PrintedAccuracy
HOG+SVM	67.24	61.55
GMM+Viterbi	72.90	88.65

Table 2. Experimental test results based on frame features and Viterbi decoder%

Method	Handwrittenaccuracy /%	Printedaccuracy /%	Frame /s
GMM+Viterbi	72.90	88.65	502
GMM+Viterbi+post-processing	78.04	89.12	496
BiLSTM	79.28	89.91	39

Table 3. Experimental results based on frame features and Viterbi decoding followed by post-processing

Method	Handwrittenaccuracy		Printedaccuracy
Method	Sent	Word	Sent	Word
Artificialsegmentation	64.92	73.01	84.67	92.10
GMM+Viterbi+post-processing	61.02	69.18	82.31	90.56
HOG+SVM	57.85	66.43	79.62	87.95

Table 4. Character recognition accuracy of different discrimination methods of handwritten and printed texts%

Scene	HOG+SVM		GMM+Viterbi+post-processing
Scene	Handwritten	Printed	Handwritten	Printed
Signed document	67.24	61.55	78.04	89.12
Natural scene	63.81	57.49	76.32	86.71
Table	65.29	57.43	72.66	86.36
Noisy document	60.31	55.23	71.48	82.23

Table 5. Classification accuracy of handwritten and printed texts after post-processing in each scene%

Scene	HOG+SVM				GMM+Viterbi+post-processing
	Handwritten		Printed		Handwritten		Printed
	Sent	Word	Sent	Word	Sent	Word	Sent	Word
Signed document	57.85	66.43	79.62	87.95	61.02	69.18	82.31	90.56
Natural scene	53.05	60.92	72.29	78.72	55.59	64.96	78.44	82.86
Table	54.61	61.98	73.89	78.73	55.16	65.01	78.60	85.21
Noisy document	45.35	54.87	66.40	72.56	48.21	56.52	68.73	76.67

Table 6. Character recognition accuracy of handwritten and printed texts in different scenes%

Qin Lin, Junfeng Xia, Zhengzheng Tu, Yutang Guo. Discrimination of Handwritten and Printed Texts Based on Frame Features and Viterbi Decoder[J]. Laser & Optoelectronics Progress, 2019, 56(6): 061003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information